Huang confirms Vera Rubin in full production with Korean memory

Nvidia Chief Executive Jensen Huang said Monday that the company's next AI accelerator, Vera Rubin, was in full production, confirming at his GTC Taipei keynote that high-bandwidth memory from Samsung Electronics, SK hynix and Micron will go into the platform.

The three suppliers were already expected. What the keynote, held alongside the Computex trade show in Taipei, settled is that the accelerator has moved to volume production, which fixes the order each maker has actually won.

That order leaves SK hynix in front. Industry estimates put it's share of HBM4 volume for the Rubin launch at roughly 60 to 70 percent, a position it built supplying the current generation. Samsung's share is estimated at 25 to 30 percent, with Micron taking a smaller, supplementary share.

For Samsung, that is still ground regained. After falling behind SK hynix on HBM, it cleared Nvidia's qualification tests at the 10 and 11 gigabit-per-second data rates the Rubin design demands and began shipping its 12-layer HBM4 early this year. Securing a quarter or more of the order on the most important AI platform marks a recovery from the rounds it lost.

HBM, the stacked memory that feeds data to AI processors, is the chip type Korea's two chipmakers compete hardest to sell, and the money at stake keeps climbing. Morgan Stanley estimates that memory now accounts for a share of the parts cost of a Rubin rack that is 435 percent higher than in the previous generation.

The contest is already moving on. Samsung has begun shipping early samples of HBM4E, the seventh-generation memory aimed at the Vera Rubin Ultra due in late 2027.

SK Group Chairman Chey Tae-won watched the keynote in person alongside SK hynix Chief Executive Kwak Noh-jung, his second appearance at an Nvidia event this year after the company's March conference in San Jose.

The keynote also reached beyond memory to other parts of SK Group. Huang showcased a digital twin of an SK hynix chip plant built with Nvidia's Omniverse simulation software, a project run with SK Telecom as part of the chipmaker's drive toward a fully autonomous fab by 2030. A digital twin recreates a real factory in virtual space so that layout and process changes can be tested before they are made on the floor.

Nvidia's reach into Korea may widen further into hardware. After the keynote, company executives told Reuters that Nvidia planned to work with robot makers in the US Europe and South Korea, on top of an existing tie-up with China's Unitree, to build humanoid robots for researchers. They did not name the partners, saying the plans were not yet public.

Huang travels to Seoul this week to meet the heads of Korea's largest conglomerates, including Chey, where the talks are expected to extend into robotics and physical AI.